FedeCouple: Fine-Grained Balancing of Global-Generalization and Local-Adaptability in Federated Learning

TL;DR

FedeCouple tackles non-IID, privacy-preserving federated learning by decoupling feature extractors and classifiers and balancing global generalization with local adaptability. It introduces integrated global-local feature learning (GLF), center loss via global feature anchors (GFA), and dynamic knowledge distillation (GPC), aided by privacy-preserving anchors and a fair, similarity-based global aggregation scheme. The authors provide a convergence analysis for nonconvex objectives and demonstrate consistent, sizable gains over nine baselines across five image datasets, including a notable 4.3% improvement in challenging settings. The approach also shows favorable communication efficiency, privacy protection, and scalability to large models, making it practical for real-world mobile and edge deployments.

Abstract

In privacy-preserving mobile network transmission scenarios with heterogeneous client data, personalized federated learning methods that decouple feature extractors and classifiers have demonstrated notable advantages in enhancing learning capability. However, many existing approaches primarily focus on feature space consistency and classification personalization during local training, often neglecting the local adaptability of the extractor and the global generalization of the classifier. This oversight results in insufficient coordination and weak coupling between the components, ultimately degrading the overall model performance. To address this challenge, we propose FedeCouple, a federated learning method that balances global generalization and local adaptability at a fine-grained level. Our approach jointly learns global and local feature representations while employing dynamic knowledge distillation to enhance the generalization of personalized classifiers. We further introduce anchors to refine the feature space; their strict locality and non-transmission inherently preserve privacy and reduce communication overhead. Furthermore, we provide a theoretical analysis proving that FedeCouple converges for nonconvex objectives, with iterates approaching a stationary point as the number of communication rounds increases. Extensive experiments conducted on five image-classification datasets demonstrate that FedeCouple consistently outperforms nine baseline methods in effectiveness, stability, scalability, and security. Notably, in experiments evaluating effectiveness, FedeCouple surpasses the best baseline by a significant margin of 4.3%.

Figures (14)

• Figure 1: Test accuracy of FedAvgref9 and FedRepref39 on CIFAR-10ref53, where smaller $\beta$ indicates stronger heterogeneity. (Left: framework, right: results)
• Figure 2: Performance evaluation on the Fashion-MNIST dataset ref55 under varying levels of heterogeneity, where a smaller value of $s$ indicates stronger heterogeneity. Tests are conducted under four conditions: 1) local training only, 2) feature extractor guided by global information ($\ell_{FE,MSE}$), 3) both feature extractor and classifier guided ($\ell_{FE,MSE}\;\&\;\ell_{CL,MSE}$), and 4) both optimally guided ($\ell_{FE,MSE}^{*}\;\&\;\ell_{CL,MSE}^{*}$). (Left: framework, right: results)
• Figure 3: Workflow of the FedeCouple method. In this process, the local feature extractor is combined with both the global frozen classifier and the local training classifier, enabling global-local feature integrated learning (GLF). Simultaneously, the global feature anchors (GFA), generated by the global feature extractor, are used to constrain the model training and construct the center loss. For the local classifier, knowledge distillation is applied to transfer knowledge from the global classifier to the local classifier. This enables the local classifier to simultaneously learn both generalization and personalization in its classification capabilities (GPC). To achieve fairer aggregation, the similarity between local models and the global average model is calculated and used as the new aggregation weight. Additionally, to fully utilize the data, data augmentation is performed using the RandAugment technique.
• Figure 4: Feature attention regions of the final convolutional layer on CINIC-10ref54 by Grad-CAM ref62. This visualization is performed in two scenarios: First, by combining the feature extractor with the global classifier (w/ Global Classifier); second, by pairing the feature extractor with the personalized local classifier (w/ Local Classifier).
• Figure 5: Representations visualization on CIFAR-10ref53 by t-SNEref69 ($\beta = 5$). The left panel shows the feature space obtained with FedRepref39, while the right illustrates the feature space after applying center loss with global feature anchors.

Theorems & Definitions (5)

• Lemma 1
• Lemma 2
• Lemma 3
• Theorem 1
• proof